cmd/internal/obj/ppc64: on Power10, use xxspltidp for float constants

Any normal float32 constant can be generated by this instruction;
use xxspltidp when possible. This prefixed instruction is much
faster than the two instruction load sequence from the
float32/float64 constant pool.

Change-Id: Id751d9ffdae71463adbde66427b986f0b2ef74c2
Reviewed-on: https://go-review.googlesource.com/c/go/+/575555
Reviewed-by: Than McIntosh <thanm@google.com>
Reviewed-by: Dmitri Shuralyov <dmitshur@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Paul Murphy <murp@ibm.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Lynn Boger <laboger@linux.vnet.ibm.com>

3 files changed, 60 insertions, 3 deletions

diff --git a/src/cmd/internal/obj/ppc64/obj9.go b/src/cmd/internal/obj/ppc64/obj9.go
index 5b7612429e..a55c402c44 100644
--- a/src/cmd/internal/obj/ppc64/obj9.go
+++ b/src/cmd/internal/obj/ppc64/obj9.go
@@ -37,6 +37,7 @@ import (
 	"internal/abi"
 	"internal/buildcfg"
 	"log"
+	"math"
 	"math/bits"
 	"strings"
 )
@@ -90,6 +91,29 @@ func isNOTOCfunc(name string) bool {
 	}
 }
 
+// Try converting FMOVD/FMOVS to XXSPLTIDP. If it is converted,
+// return true.
+func convertFMOVtoXXSPLTIDP(p *obj.Prog) bool {
+	if p.From.Type != obj.TYPE_FCONST || buildcfg.GOPPC64 < 10 {
+		return false
+	}
+	v := p.From.Val.(float64)
+	if float64(float32(v)) != v {
+		return false
+	}
+	// Secondly, is this value a normal value?
+	ival := int64(math.Float32bits(float32(v)))
+	isDenorm := ival&0x7F800000 == 0 && ival&0x007FFFFF != 0
+	if !isDenorm {
+		p.As = AXXSPLTIDP
+		p.From.Type = obj.TYPE_CONST
+		p.From.Offset = ival
+		// Convert REG_Fx into equivalent REG_VSx
+		p.To.Reg = REG_VS0 + (p.To.Reg & 31)
+	}
+	return !isDenorm
+}
+
 func progedit(ctxt *obj.Link, p *obj.Prog, newprog obj.ProgAlloc) {
 	p.From.Class = 0
 	p.To.Class = 0
@@ -111,7 +135,7 @@ func progedit(ctxt *obj.Link, p *obj.Prog, newprog obj.ProgAlloc) {
 	// Rewrite float constants to values stored in memory.
 	switch p.As {
 	case AFMOVS:
-		if p.From.Type == obj.TYPE_FCONST {
+		if p.From.Type == obj.TYPE_FCONST && !convertFMOVtoXXSPLTIDP(p) {
 			f32 := float32(p.From.Val.(float64))
 			p.From.Type = obj.TYPE_MEM
 			p.From.Sym = ctxt.Float32Sym(f32)
@@ -123,7 +147,7 @@ func progedit(ctxt *obj.Link, p *obj.Prog, newprog obj.ProgAlloc) {
 		if p.From.Type == obj.TYPE_FCONST {
 			f64 := p.From.Val.(float64)
 			// Constant not needed in memory for float +/- 0
-			if f64 != 0 {
+			if f64 != 0 && !convertFMOVtoXXSPLTIDP(p) {
 				p.From.Type = obj.TYPE_MEM
 				p.From.Sym = ctxt.Float64Sym(f64)
 				p.From.Name = obj.NAME_EXTERN
diff --git a/test/codegen/floats.go b/test/codegen/floats.go
index d5c5475567..baa745bdee 100644
--- a/test/codegen/floats.go
+++ b/test/codegen/floats.go
@@ -196,3 +196,34 @@ func Float32Max(a, b float32) float32 {
 	// ppc64/power10:"XSMAXJDP"
 	return max(a, b)
 }
+
+// ------------------------ //
+//  Constant Optimizations  //
+// ------------------------ //
+
+func Float32Constant() float32 {
+	// ppc64x/power8:"FMOVS\t[$]f32\\.42440000\\(SB\\)"
+	// ppc64x/power9:"FMOVS\t[$]f32\\.42440000\\(SB\\)"
+	// ppc64x/power10:"XXSPLTIDP\t[$]1111752704,"
+	return 49.0
+}
+
+func Float64Constant() float64 {
+	// ppc64x/power8:"FMOVD\t[$]f64\\.4048800000000000\\(SB\\)"
+	// ppc64x/power9:"FMOVD\t[$]f64\\.4048800000000000\\(SB\\)"
+	// ppc64x/power10:"XXSPLTIDP\t[$]1111752704,"
+	return 49.0
+}
+
+func Float32DenormalConstant() float32 {
+	// ppc64x:"FMOVS\t[$]f32\\.00400000\\(SB\\)"
+	return 0x1p-127
+}
+
+// A float64 constant which can be exactly represented as a
+// denormal float32 value. On ppc64x, denormal values cannot
+// be used with XXSPLTIDP.
+func Float64DenormalFloat32Constant() float64 {
+	// ppc64x:"FMOVD\t[$]f64\\.3800000000000000\\(SB\\)"
+	return 0x1p-127
+}
diff --git a/test/codegen/math.go b/test/codegen/math.go
index 331ebbe609..eb6e927dec 100644
--- a/test/codegen/math.go
+++ b/test/codegen/math.go
@@ -204,7 +204,9 @@ func constantCheck32() bool {
 func constantConvert32(x float32) float32 {
 	// amd64:"MOVSS\t[$]f32.3f800000\\(SB\\)"
 	// s390x:"FMOVS\t[$]f32.3f800000\\(SB\\)"
-	// ppc64x:"FMOVS\t[$]f32.3f800000\\(SB\\)"
+	// ppc64x/power8:"FMOVS\t[$]f32.3f800000\\(SB\\)"
+	// ppc64x/power9:"FMOVS\t[$]f32.3f800000\\(SB\\)"
+	// ppc64x/power10:"XXSPLTIDP\t[$]1065353216, VS0"
 	// arm64:"FMOVS\t[$]\\(1.0\\)"
 	if x > math.Float32frombits(0x3f800000) {
 		return -x